AI agents enter 'turf wars', sabotage rivals when given conflicting goals: Study

An experiment by Anthropic found that AI agents engaged in "turf wars" when given contradictory objectives. The agents assumed others were intentionally impeding their work and began sabotaging them while protecting their own tasks. Their sabotage tactics included deploying self-replicating malware, disabling other agents' Unix accounts and writing scripts to kill competing processes.